Nobody knows which geospatial foundation model is best.
CROMA scores 30.1 mIoU on m-Cashew in one paper and 62.2 in another. Of the 252 results that more than one paper reports under the same protocol, 47 differ by 5 points or more.
Lowest and highest score reported for the same experimentlowest highest
020406080100
Paper, data, and slides
research.taylorgeospatial.org/state-of-gfms
We have published tables like these ourselves, and we think fixing them takes shared benchmarks and tooling that no single lab can provide.